The Soulful Voice Of Dua Lips Music As Dua Where Desire Meets Divinity

Working With Vocal Synthesis Models That Emulate Pop Vocals

I spent roughly six months testing various voice synthesis approaches for a commercial project last year. The goal was to create backing vocal textures that sounded convincingly like a specific contemporary pop vocalist without licensing an actual artist. This meant working closely with what I'll refer to throughout as The Soulful Voice Of Dua Lips Music As Dua Where Desire Meets Divinity — a fine-tuned RVC (Retrieval-based Voice Conversion) model built on a substantial dataset of her recorded output. Here is how it actually functions in practice. The model operates on RVC architecture, which separates timbre from content. You feed it a dry vocal source — any singing voice — and it maps the spectral characteristics of the target onto your input. The result is not a perfect clone. It is an approximation that captures pitch contours, vibrato patterns, and certain resonance qualities while preserving the original performance's emotion and phrasing. The training dataset matters enormously. Most publicly available versions I encountered were trained on around 15 to 20 minutes of cleaned vocal stems. That is the bare minimum. The better models use 40 minutes or more, carefully filtered for consistent mic technique and minimal reverb. I learned this the hard way after wasting three days on a model that sounded thin and nasal on high notes because the training data was pulled from live performances rather than studio recordings.

Setting Up The Pipeline

I run everything locally on a machine with an NVIDIA RTX 3090, 24GB VRAM. The standard approach uses RVC WebUI, which handles inference. You need three components: the base model (typically a RMVX or Demucs stem separator to isolate vocals), the RVC inference engine, and the voice model file itself. Here is the workflow I settled on after trial and error: First, separate your source vocal from any instrumental using Demucs v4 with the "htdemucs" variant. The "mdx" preset sometimes bleeds synths into the vocal track, which introduces artifacts after conversion. Set the output to 44.1kHz WAV. Then load the voice model into RVC WebUI and adjust the pitch shift. For a female-to-female conversion, a +0 semitone shift preserves the natural range. Going +12 makes it sound artificial and strained around the upper register.

The most critical setting is the reverb mix. By default, RVC outputs a completely dry signal. The original Dua recordings have very specific spatial characteristics — a tight plate reverb around 1.2 seconds decay with a pre-delay of about 25 milliseconds. If you do not add matching spatial processing, the converted vocal will sit oddly in the mix, clearly identifiable as synthetic.

A Specific Problem And What Worked

About four months into this project, I hit a wall. The model would handle sustained notes beautifully, but any fast melisma or rapid phrasing — think the verses of "Levitating" or "Physical" — produced noticeable robotic artifacts. The pitch detection would skip between semitones instead of sliding smoothly. The workaround involved preprocessing. I ran the source vocal through a gentle pitch correction pass using Melodyne's "Graphical" mode, just enough to give the pitch estimator a cleaner target. Not full correction, just nudging wild pitches toward their intended notes. Then I increased the f0 method to "crepe-full" in the RVC settings. This is slower — roughly three times the real-time factor — but it produces dramatically smoother pitch contours on fast passages. Combined with a threshold of 0.3 for index-based refinement, the artifacts dropped to nearly invisible levels. The tradeoff is render time. A two-minute vocal take that previously converted in eight minutes now takes twenty-four. Worth it for final export. Not worth it for rapid iteration.

Common Pitfalls Beginners Miss

The biggest mistake is assuming the model transfers cleanly across genres. This voice model was trained primarily on dance-pop and disco-revival material. When you run a folk-style acoustic guitar vocal through it, the model forces pop vocal techniques onto material that does not need them. The result sounds wrong in a subtle, unmistakable way. The breath patterns become too controlled, the tonal brightness pushed too far forward. Another issue people overlook: the model tends to add a slight nasality to low-register passages. The training data had fewer low notes relative to mid-range content. If your source vocal dips below G3, you will hear a nasal resonance that was never there. The fix is straightforward — apply a gentle EQ cut around 800Hz to 1.2kHz on the converted output, but only on those lower passages. Automate the EQ rather than applying it globally.

Limitations You Should Know About

This approach does not solve every problem. Here is what it cannot do reliably: It struggles with extreme dynamic contrasts. A whisper-quiet intro followed by a belted chorus in the same take will produce inconsistent timbre between sections. The model averages characteristics across the input. For projects requiring wide dynamic range, you need to split the vocal into sections and process each separately. Emotional authenticity remains the unresolved frontier. The model captures technical qualities — vibrato speed, breathiness ratio, formant distribution — but it cannot replicate the micro-decisions a human singer makes under emotional pressure. Slight pitch imperfections that convey vulnerability get smoothed out. The voice sounds technically correct and emotionally flat.

If your project requires genuine emotional nuance over extended passages, the better path is hiring a session vocalist and using these models only for doubling or textural purposes. That is where they genuinely shine.

Getting Started

The RVC community maintains GitHub repositories with the base software. Training your own model from scratch requires significant GPU time — roughly 6 to 12 hours on a 3090 for a adequate dataset — but pre-trained models are available through community sharing platforms. I typically evaluate a model by running a 30-second test phrase through it and checking for three things: timbre accuracy on sustained notes, artifact presence on rapid passages, and naturalness of breath sounds between phrases. If two out of three look good, the model is usable. If none do, move on rather than spending hours trying to tweak parameters that cannot fix a poorly trained base. For the conversion itself, RVC WebUI version 2.x remains the most stable release. Pair it with Demucs v4 for stem separation and a quality UVR5 implementation if you need more control over the isolation stage. The total setup time from clean vocal input to mixed output averages about forty-five minutes per two-minute track at the quality level I described above.